Article ID: 22.20250364
The integration of third-party IP cores into IC designs significantly increases the risk of stealthy hardware Trojans (HTs). Existing detection approaches, based on functional testing, logic verification, or machine learning, often focus on individual gates and overlook the structural patterns across Trojan circuits. Moreover, they lack diagnostic capabilities and typically require manual inspection. This work presents TrojanHound, a hierarchical detection framework that unifies gate-level analysis and circuit-level diagnosis. It leverages a graph neural network with adaptive neighborhood aggregation to identify suspicious nodes, merges related components into candidate subcircuits using subgraph fusion, and verifies HT presence through topology-aware analysis based on betweenness centrality. Experiments on 14 TrustHub benchmarks show that TrojanHound achieves a 95.64% true positive rate and a 97.44% F1-score, without false positives. By integrating structural learning with topological reasoning, this method enables accurate and automated HT detection in complex ICs.
